INFINITY on AirfRANS

Reproduction of INFINITY: Neural Field Modeling for Reynolds-Averaged Navier-Stokes Equations (arXiv:2307.13538) on the AirfRANS dataset. Implementation is derived solely from the paper text (Figure 1, Algorithm 1, Table 1) and the provided data.

Model

• 6 modulated INRs (Fourier Features + FiLM shift): d, n, vx, vy, p, nut
• Hypernetworks mapping latent code z_u -> modulation phi_u
• Mapping network g_psi (MLP): (z_d, z_n, Vx, Vy) -> (z_vx, z_vy, z_p, z_nut)
• Two-stage training: (1) CAVIA second-order meta-learning of INRs (Algorithm 1, K=3 inner steps), (2) supervised MSE in z-code space for g_psi.

Structure

• config/ - Configuration YAML and manifest JSON files
• model/ - Model architecture definitions (INFINITY, ModulatedINR, FourierFeatures, Hypernetwork)
• scripts/ - Training, evaluation, data loading, and utility scripts
• weight/ - Pre-trained model checkpoint files for various configurations

Run

python scripts/train.py config/infinity.yaml        # stage 1 + stage 2
python scripts/evaluate.py config/infinity.yaml     # Table-1 style metrics

Config

All hyperparameters in config/infinity.yaml (paper only fixes K=3; others are default assumptions).

Configurations